Code for the "Compiling ruby" talk
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
bin
sample
tuby @ 23a6374
vernacular @ fb43635
.gitignore
.gitmodules
Gemfile
LICENSE
README.md
Rakefile

README.md

Compiling Ruby

Since Ruby 2.3 we've been able to programmatically load ruby bytecode through the introduction of RubyVM::InstructionSequence::load_iseq. If the ::load_iseq method is defined, every file that is required will pass through it. If ::load_iseq then returns nil, the normal process of parsing the Ruby file will still happen. Otherwise, it should return a precompiled set of Ruby instructions through an instance of RubyVM::InstructionSequence. To read more about the initial proposal for this feature, see the ruby issue.

By divorcing the process of running YARV bytecode from the process of compiling Ruby code, Ruby can improve boot speed (because the Ruby source no longer has to be parsed into instruction sequences) and also reduce memory consumption on boot. These were both of the stated goals of ::load_iseq, and it is accomplished and demonstrated in both the yomikomu and bootsnap community gems.

For an example, let's take a look at a simple implementation of ::load_iseq:

class RubyVM::InstructionSequence
  ISEQ_DIR = File.expand_path(__dir__, '.iseq')

  def self.load_iseq(source_path)
    iseq_name = source_path.gsub(/[^A-Za-z0-9\._-]/) { |c| '%02x' % c.ord }
                           .gsub('.rb', '.yarb'))
    iseq_path = File.join(ISEQ_DIR, iseq_name)

    if File.exist?(iseq_path) && (File.mtime(source_path) <= File.mtime(iseq_path))
      return RubyVM::InstructionSequence.load_from_binary(File.binread(iseq_path))
    end

    begin
      contents = File.read(source_path)
      iseq = RubyVM::InstructionSequence.compile(contents)
      File.binwrite(iseq_path, iseq.to_binary)
      iseq
    rescue SyntaxError, RuntimeError
      nil
    end
  end
end

For the above implementation, when ::load_iseq is called, it will check to determine whether or not a compiled instruction sequence file has already been created for the given source file. If it has, and it's up to date with the latest source, it will return the compiled instruction sequence. If it is not, it will go through the process of compiling the source into bytecode, write that out for future runs, and then return the generated instruction sequences.

Preprocessors

The above implementation works well and accomplishes the stated goals of the issue. However, there are other possibilities that this new functionality has enabled. After the source is read but before it is compiled, the contents can be preprocessed in any way you want.

Modifying as text

As an example, we can modify the text in place without knowledge of the lexical context, as in Elixir-like sigils:

contents = File.read(source_path)

# implement date sigils
format = '%FT%T%:z'
contents.gsub!(/~d\((.+?)\)/) do |match|
  content = match[3..-2]
  begin
    date = Date.parse(content)
    "Date.strptime('#{date.strftime(format)}', '#{format}')"
  rescue ArgumentError
    "Date.parse(#{content})"
  end
end

# implement number sigils
contents.gsub!(%r{~n\(([\d\s+-/*\(\)]+?)\)}) { |match| eval(match[3..-2]) }

# implement URI sigils
contents.gsub!(/~u\((.+?)\)/, 'URI.parse("\1")')

iseq = RubyVM::InstructionSequence.compile(contents)

This allows us clearly express date, numbers, and URIs within our code. It also has a couple of other benefits:

  • for literal dates in the date sigil, it allows the ~d(January 1st, 2017) format. This would normally be inefficiently parsed with Date::parse, but because of preprocessing actually gets parsed using the much more efficient Date::strptime
  • for number sigils, the entire contents of the sigil is evaluated before the code is compiled. This means less memory being allocated and fewer instructions being sent to the VM. For example, usually 24 * 60 * 60 would require two function calls and three stack-allocated numbers. With a sigil (~n(24 * 60 * 60)) it's just one stack-allocated number and no function calls.

Modifying as an AST

The other modification we can perform is after the text has been parsed into an abstract syntax tree (and we therefore have lexical context). For example, we can use the parser gem to build our own AST that contains invalid Ruby, rewrite it to a valid Ruby AST, and continue on, as in:

content = File.read(source_path)

require 'parser/current'

# class CustomBuilder < Parser::Builders::Default
builder  = CustomBuilder.new

# class CustomParser < Parser::CurrentRuby
parser   = CustomParser.new(builder)

# class CustomRewriter < Parser::Rewriter
rewriter = CustomRewriter.new
rewriter.instance_variable_set(:@parser, parser)

buffer = Parser::Base.send(:setup_source_buffer, '(string)', 1,
                           content, parser.default_encoding)
ast = parser.parse(content)
content = rewriter.rewrite(buffer, ast)

iseq = RubyVM::InstructionSequence.compile(content)

Vernacular

All of these modifications and more are encompassed by the vernacular gem, included in this repository.